A truthfulness audit takes every claim on a CV, separates it from the document, and asks one question of each: what evidence backs this? The answer is one of three verdicts — supported, unsupported, or contradicted. The middle one is the one people misread. Unsupported does not mean false; it usually means you left the evidence out.
What is a resume truthfulness audit?
A resume truthfulness audit is a systematic check of every claim in the document against the evidence behind it — the same audit whether you call the document a resume or a CV; this page says CV. Each assertion — a skill, an achievement, a metric, a job title, a leadership or seniority claim — is pulled out of the document as a standalone statement, then tested against what your actual career record supports. Each claim receives one of three verdicts: supported, meaning the record backs it directly or follows safely from it; unsupported, meaning nothing in the record backs it up; or contradicted, meaning it conflicts with something the record already says.
The audit is deliberately separate from the writing. Whoever wrote the CV does not get to grade it, because the writer has an interest in the answer. And the output is not a verdict on you as a candidate. It is a list of the specific lines you would struggle to defend if an interviewer said “walk me through that one.”
The three verdicts, and what each one actually means
| Verdict | What it means | What to do about it | What happens if you ignore it |
|---|---|---|---|
| Supported | Traces to something concrete in your record — stated directly, or safely derived from what is stated. | Nothing. Leave it. If it’s a strong claim, move it higher up the page. | Nothing. This is the pile you want. |
| Unsupported | Nothing in your record backs it up yet. Not a lie — an unevidenced assertion. | Add the evidence you compressed out, or soften the claim to what you can defend. | You get asked, produce nothing specific, and the interviewer quietly discounts the rest of the CV too. |
| Contradicted | Conflicts with something else on your own CV — dates, tenure, title, scope. | Fix it now. One of the two statements is wrong; work out which. | Reads as dishonesty rather than sloppiness, and it’s visible to anyone who reads the document carefully. |
The supported pile splits in practice: some claims are backed explicitly by something written down, others safely derived from it — “worked in a regulated environment”, derived from three years at a bank. Derived is legitimate but weaker, and should never be promoted to explicit because it reads better.
Why the unsupported/contradicted distinction is the useful part
Most CV advice collapses everything into “don’t lie” — true, and useless, because almost nobody writing a CV thinks of themselves as lying. The two failures are different problems with different fixes.
Contradicted is an error. “Five years of Python” when your employment history covers two. “Led the team” in one bullet, “supported the team lead” in another. That’s a defect in the document; fix it by deciding what’s true and making the CV say that.
Unsupported is usually an omission. You wrote “improved deployment reliability” and nothing else on the page mentions deployments, incidents or reliability. The claim is probably true — you lived it, then compressed it to fit two pages, and everything that made it credible fell out. That’s why it’s the productive verdict: not an accusation, a prompt. The evidence exists in your head and not on the page. Most unsupported claims are fixed by adding, not deleting.
How the gating works
An audit that only produces a list is easy to ignore, so in SiviGen the verdicts drive fixed rules:
- A contradicted claim marks the whole result blocked. Not a warning, not a deduction — the result carries a not ready to send verdict and names the two statements that conflict. Be clear about what that is and isn’t: it is a verdict, not a lock. Nothing stops you exporting the PDF. What it stops is you exporting it without having been told.
- A major unsupported claim raises a warning. “Major” means the claim types that hurt most when they turn out hollow: achievements, metrics, seniority, leadership, titles, certifications. An unsupported soft skill is minor. An unsupported certification is not.
- A truthfulness score below 85 caps the overall score at 59. However relevant the CV is, it can’t present as a good result while the honesty score is weak. The cap stops the two numbers being traded against each other.
The score starts at 100 and loses points per problem claim, weighted by type, each claim contributing at most one penalty — its worst applicable one. The verdicts, the gates and the scoring are the last stage of the tailoring pipeline behind the claims. None of it is specific to a CV, either: a cover letter goes through the same extractor and verifier, not a gentler version of one.
Why the score is computed in code, not asked of a model
The model’s job is to produce structured facts: here is a claim, its type, whether the record supports it, how severe the gap is. Ordinary TypeScript computes the numbers and applies the gates.
That isn’t an optimisation, it’s the whole point. A model asked “how truthful is this CV you just wrote?” is marking its own homework, and it does what anyone marking their own homework does. Splitting the work means the writer never touches the number, the same set of claim verdicts always produces the same score, and every deduction traces to a named claim and a readable rule.
I should be precise about what that buys, because it is easy to oversell. The arithmetic is fixed; its inputs are not. Which claims get extracted, and whether your record is judged to support each one, come from a model, and two runs over the same CV can differ there. What the split guarantees is that the document’s author never gets to grade the document, and that every point lost is attributable to a claim you can read. The same principle governs the interview readiness score: a model observes, code decides.
How to audit your own CV by hand
No account needed. You need your CV, a pen, and about thirty minutes for a two-page document.
1. Break it into claims. Go through the CV one line at a time. Each bullet usually holds one to three separate assertions; write them out as standalone statements. “Led migration of the billing system to AWS, cutting costs 40%” is three claims: you led it, it was a migration to AWS, it cut costs by 40%.
2. Name the evidence. For each claim, ask the interviewer’s question: if someone said “walk me through that”, what specifically would I say? Name the artefact — the project, the incident, the number, the person who’d confirm it. Give yourself ten seconds.
3. Sort into three piles. Evidence arrived immediately: supported. You hedged, generalised, or had to think: unsupported. Something else on your CV says otherwise: contradicted.
4. Fix contradicted first. Check dates against tenure claims, titles against scope claims, “led” against everything. Quick, and the most damaging.
5. For each unsupported claim, add or soften. Add where the evidence exists. Soften where it doesn’t.
A worked example
The bullet: “Improved system reliability and reduced downtime by 40%.”
What would you produce if challenged? You remember the on-call rotation was miserable, you added retries and a circuit breaker to the payments callout, and the pager stopped going off. But 40%? You don’t know where that came from — you wrote it because a number looked stronger than no number. Verdict: unsupported, high severity, because unsupported metrics are the fastest way to lose a room. Two honest fixes:
- Add the evidence. You check: incidents in that service went from roughly weekly to roughly monthly in the quarter after the change. Rewrite as “Cut recurring payment-callout incidents from ~weekly to ~monthly by adding retries and a circuit breaker to the upstream integration.” Longer, specific, and you can walk through every word.
- Soften the claim. If you can’t reconstruct the numbers, drop them: “Reduced repeat incidents in the payments integration by adding retries and circuit-breaking to the upstream callout.” No metric, still concrete, still defensible.
Both move the claim to supported. The first is stronger. Neither invents anything.
To pressure-test one line before rewriting the whole document, the resume claim checker flags the claims most likely to need evidence. It works from your wording alone — it never sees your history, so step 2 stays yours — and it runs in the browser with no signup. When the problem is step 5 and you know you did something good but can’t reconstruct the detail, the achievement finder asks the questions that find the evidence you’re missing. And since the honest fix for an unsupported claim is usually more words, not fewer, the resume length checker tells you how much room you’re actually working with before you start adding.
What a hand audit won’t catch
Drift. You audit the master CV once, then tailor it eleven times in three weeks, and each version quietly grows a claim that suited that particular job description. The audit has to run on the version you actually send, not the one you cleaned up in January. That’s the case for automating it, and it’s what SiviGen does — but the method above is the whole method. Run it by hand if you like. The point is that it gets run.
If you are choosing a tool to run it for you, the question worth asking is not whether the vendor promises accuracy but what the tool does when the job asks for something your history doesn’t contain. That takes ten minutes to answer yourself: hand it a job description demanding a skill your CV never mentions, and read what it writes back. Any tool worth using — mine included — should leave the gap visible rather than fill it.
FAQ
Is an unsupported claim the same as a lie?
No. Unsupported means nothing on your CV currently backs the claim up — the evidence may well exist, it just isn’t on the page. Most unsupported claims are true things you compressed too far, and the fix is to add the missing specifics rather than delete the line. Contradicted is the serious category: that one conflicts with something your CV already says.
What counts as a contradicted claim?
Anything that conflicts with your own document: more years of a technology than your employment dates cover, “led the team” in one bullet and “supported the team lead” in another, a certification dated after the role where you claim you used it. These are the claims a careful reader catches without asking you a single question.
Does auditing my CV make it weaker?
Usually the opposite. Removing an unverifiable number costs you one vague digit; replacing it with a specific, defensible detail gains you a story you can tell out loud. The bullets that survive an audit are the ones you’ll answer confidently when asked.
Can an AI tool audit claims reliably?
Only if the part that writes is not the part that judges. A single model asked to write a CV and then rate its own honesty will rate it well. The design that works separates the two: extract the claims from the finished draft as if seeing it for the first time, verify each against a fixed record of what the candidate did, then compute the score in code.
If you’d rather this ran on every tailored version than once by hand: SiviGen tailors your CV from a fact graph of your real experience, audits every claim on the result, and marks the CV not ready to send when a claim contradicts your own record — naming the two statements that conflict, so there is something specific to fix. It will not stop you downloading the PDF; it will stop you doing it unaware, which is the part that costs you an interview. The rest of what SiviGen does is built on the same split: the model writes, code decides. It’s free.